Claude Agent with Dynamic Skills on Amazon Bedrock AgentCore

---

This sample demonstrates how to deploy an AI agent using the Claude SDK with Amazon Bedrock on Amazon Bedrock AgentCore Runtime, with dynamic skill definitions loaded from Amazon S3.

What You'll Learn

• Deploy a containerized agent to AgentCore Runtime using Claude SDK
• Load skill definitions dynamically from S3 at runtime
• Configure IAM permissions for agent execution
• Use --local-build to avoid Lambda layer injection issues

Problem

Customers needs Claude Agent SDK's skill system on AgentCore Runtime, but a mismatch between Claude SDK's filesystem-based approach (expecting .claude/skills/ directories with complex nested structures, Python scripts, and interdependent skill.md files) and AgentCore's constraints that limit filesystem access and container size. With an ever-growing skill repository (10-30 skills expanding to 100+) that needs to be shared across multiple agent runtimes, the standard approach of bundling all skills into each container becomes unmanageable and prevents hot updates. Our S3-based solution provides centralized skill storage, dynamic runtime loading, and hot-swappable capabilities essentially solving enterprise-scale skill management challenges that the native Claude Agent SDK can't address on constrained runtime environments.

Architecture

┌──────────┐     ┌─────────────────────┐     ┌─────────────────┐
│  User    │────▶│  AgentCore Runtime  │────▶│   Container     │
│  Request │     │  (Managed Service)  │     │   (HTTP:8080)   │
└──────────┘     └─────────────────────┘     └────────┬────────┘
                                                      │
                         ┌────────────────────────────┼────────────┐
                         │                            ▼            │
                         │  ┌─────────┐       ┌─────────────┐     │
                         │  │   S3    │◀──────│  Claude SDK │     │
                         │  │ Skills  │ load  │  + Bedrock  │     │
                         │  └─────────┘       └─────────────┘     │
                         └─────────────────────────────────────────┘

Prerequisites

• An AWS account with credentials configured (aws configure)
• Python 3.10 or later
• Docker installed and running
• Model Access: Claude model enabled in Amazon Bedrock console
• AWS Permissions:
  • BedrockAgentCoreFullAccess managed policy
  • AmazonBedrockFullAccess managed policy

Quick Start

Step 1: Clone Repository

git clone <your-repo-url>
cd Claude-Agent-Skill

Step 2: Install Dependencies

pip install bedrock-agentcore anthropic boto3

Step 3: Configure the Agent

agentcore configure -e claude_sdk_bedrock.py -dt container -dm -ni

Step 4: Deploy to AgentCore Runtime

# Important: Use --local-build to avoid Lambda layer injection
agentcore launch --local-build

Step 5: Test Your Agent

# Check status
agentcore status

# Invoke the agent
agentcore invoke '{"prompt": "What skills do you have?"}'

# Try another prompt
agentcore invoke '{"prompt": "Write a Python hello world"}'

Success: You should see responses that reference the skills loaded from S3.

S3 Skills Structure

Create an S3 bucket with skill definitions:

your-skills-bucket/
└── skills/
    ├── code-analysis/
    │   └── skill.md
    ├── web-research/
    │   └── skill.md
    ├── data-fetcher/
    │   └── skill.md
    ├── document-generator/
    │   └── skill.md
    ├── report-generator/
    │   └── skill.md
    └── data-analysis-workflow/
        └── skill.md

Each skill.md file contains a description of the skill's capabilities that gets injected into the system prompt.

IAM Permissions

Add S3 access to your AgentCore execution role:

{
  "Effect": "Allow",
  "Action": ["s3:GetObject", "s3:ListBucket"],
  "Resource": [
    "arn:aws:s3:::your-skills-bucket",
    "arn:aws:s3:::your-skills-bucket/*"
  ]
}

Configuration

Setting	Value
Model	us.anthropic.claude-sonnet-4-20250514-v1:0
Port	8080
Protocol	HTTP
Build	Local (--local-build)

Environment Variables

Variable	Description	Default
AWS_REGION	AWS region	us-east-1
SKILLS_S3_BUCKET	S3 bucket for skills	claude-skills-poc-bucket
ANTHROPIC_MODEL	Bedrock model ID	us.anthropic.claude-sonnet-4-20250514-v1:0

Key Implementation Details

Claude SDK Integration

import anthropic
from bedrock_agentcore.runtime import BedrockAgentCoreApp

# Claude SDK with Bedrock backend
claude_client = anthropic.AnthropicBedrock(aws_region=AWS_REGION)

app = BedrockAgentCoreApp()

@app.entrypoint
def handler(event, context):
    response = claude_client.messages.create(
        model=MODEL_ID,
        max_tokens=1024,
        system=system_prompt,
        messages=[{"role": "user", "content": prompt}]
    )
    return {"response": response.content[0].text}

Dynamic Skill Loading

Skills are loaded from S3 at container startup and injected into the system prompt:

def load_skills():
    response = s3_client.list_objects_v2(Bucket=SKILLS_S3_BUCKET, Prefix='skills/', Delimiter='/')
    for prefix in response.get('CommonPrefixes', []):
        name = prefix['Prefix'].replace('skills/', '').rstrip('/')
        desc = s3_client.get_object(Bucket=SKILLS_S3_BUCKET, Key=f'skills/{name}/skill.md')
        SKILLS[name] = desc['Body'].read().decode('utf-8')[:200]

Cleanup

agentcore destroy

Related Resources

• Amazon Bedrock AgentCore Documentation
• Amazon Bedrock AgentCore Samples
• Claude SDK Documentation
• AgentCore Starter Toolkit